Generative Adversarial Networks
Generative Adversarial Networks (GANs) consist of two competing networks: a generator (creates fake data) and a discriminator (tries to distinguish real from fake). They are trained together in a minimax game, leading the generator to produce highly realistic samples.
Generator tries to fool discriminator; discriminator tries to catch fakes. Eventually, generator produces realistic data.
How GANs Work
1. Generator takes random noise z and generates fake sample G(z).
2. Discriminator takes both real and fake samples and outputs probability of being real.
3. Generator loss: wants discriminator to be wrong (classify fake as real).
4. Discriminator loss: wants to correctly classify real vs fake.
5. Alternate training: update discriminator, then generator, repeatedly.
Challenges
- Mode collapse: generator produces only few types of samples.
- Training instability: discriminator may become too strong.
- Difficult to evaluate quality objectively.
Famous GAN Variants
- DCGAN (Deep Convolutional GAN): uses conv layers for image generation.
- StyleGAN: generates high‑quality, controllable faces.
- CycleGAN: unpaired image‑to‑image translation (e.g., horse ↔ zebra).
- Pix2Pix: paired image translation (e.g., sketch → photo).
- WGAN (Wasserstein GAN): improves training stability.
Applications
- Image generation (faces, art, objects).
- Image super‑resolution, colorization.
- Data augmentation for rare classes.
- Text‑to‑image synthesis (e.g., GANs before diffusion).
Two Minute Drill
- GANs have generator and discriminator competing.
- Generator learns to create realistic fakes.
- Challenges: mode collapse, instability.
- Used for image generation, translation, super‑resolution.
Need more clarification?
Drop us an email at career@quipoinfotech.com